Unverified50% confidenceFactExact time

RLHF训练中会加入KL散度惩罚项，防止模型偏离SFT阶段学到的基础行为太远

1

Sources

50%

Confidence

Long-term

Relevance

8/21/2026

First Seen

Sources

How Reinforcement Learning Creates AI's 'Split Personality': The Deep Flaws of RLHF Alignment

Related Claims

UnverifiedRLHF和DPO等对齐技术的效果高度依赖预训练基础质量，预训练阶段未习得的能力几乎无法通过后续对齐补救72% similar Verified当代前沿模型的训练方法包括RLHF（基于人类反馈的强化学习）及其后继者RLAIF和过程奖励模型（PRM）71% similar Unverified强化学习训练比监督学习更不稳定，将NVFP4激进量化引入RL训练会进一步放大固有不稳定性70% similar Unverified当前主流LLM普遍采用预训练加微调范式，先在互联网规模语料上无监督预训练，再通过RLHF等技术对齐人类偏好69% similar

Cite This Claim

Stable URI

https://kongchang.com/claim/784174

API

curl https://kongchang.com/api/v1/knowledge/claims/784174

MCP

get_claim(id=784174)